v0.5.4

Image bridge: settings.imageBridge (default false, opt-in) gives a text-only pro model second-hand vision. When a request carries images only in history — the final message is text — awerouter has the multimodal flash destination (imageModel) transcribe each distinct image, replaces the image blocks with the transcription text before routing, and the session falls through to defaultModel (pro) instead of being pinned to flash forever. A fresh upload this turn still routes to flash natively; router.py is unchanged — after the rewrite has_image is false and the request lands in L3/default, before the rewrite the L1 image guard routes it as always. The step-glm-mm template now ships with imageBridge: true.

Added

• settings.imageBridge (true/false, global or per profile, validated at load and shown by config show): the request-pipeline transform described above. It runs before rtk compression and routing in _proxy_flow, so what rtk compresses, what L3 scores, and what goes upstream are exactly the same body; handle_count_tokens applies the same bridge so the client's context-window estimate matches. The serve banner prints image bridge -> on (...) naming the transcribing destination.
• awerouter.vision (pure logic): per-protocol content keys (image_key), caption body building and response parsing for all three wire protocols, in-place image-block rewriting ([Image n, transcribed by <model>] text blocks), and a content-addressed caption cache (bounded FIFO, process lifetime — a restart re-transcribes each distinct image once; the cache key includes provider+model, so a destination switch never serves another model's transcription).
• Failure fallback: any caption call failure (network, non-200, empty text) leaves the request body untouched — one printed line, and the L1 image guard routes the request exactly as before. Codex subscription logins (auth: "codex", SSE-only backend) skip the bridge entirely.
• step-glm-mm template sets imageBridge: true; init --merge flags a newly-written imageBridge as a behavior shift next to imageModel/defaultModel. README/README_cn document it as an "Image bridge" (图片桥接) section; README.ai.md and the awerouter skill mention the template flag.

Behavior notes

• Bridged "vision" is second-hand: its ceiling is the transcription quality, and each distinct image costs one extra flash call (cached per image content for the process lifetime).
• The flash→pro rescue fallback (flash down on an un-rewritten image request) still carries raw images to a possibly text-only pro — pre-existing behavior, unchanged and out of scope.